[
https://issues.apache.org/jira/browse/HBASE-2925?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12905987#action_12905987
]
Robert Mahfoud commented on HBASE-2925:
---------------------------------------
@stack, thank you for picking up this issue. Feel free to reuse the code
provided in SimpleHConnectionManagerLeakReplicator.java in any way you want.
(Let me know if this is not enough I could submit another file with Apache
license added.)
Back to the discussion, I agree that removing the {{hasCode()}} and
{{equals()}} methods does resolve this issue. However, I wonder if it defeats
the purpose of having a connection cache from the first place. You also alluded
to the problem of the configuration settings changing after the connection is
open, which, if we rely solely on Object equality, could lead to subtle error
or unexpected behavior.
Hence my suggestion to pin down the configuration once the TableServers
instance is created. The disadvantage there is that small/unrelated changes in
the configuration will invalidate the cache entry as well.
If that is deemed a real problem given the common use cases, I was thinking
maybe we define "equality" of the configuration for the purposes of caching as
the equality of all properties that start with {{hbase.}} or {{zk.}} or any
other prefix whose properties could be used as possible connection parameters.
Not sure if the comparison becomes too slow in that case, but what is more
important is whether correctness is compromised if we make the assumption that
only properties starting with those prefixes affect the connection.
> LRU of HConnectionManager.HBASE_INSTANCES breaks if HBaseConfiguration is
> changed
> ---------------------------------------------------------------------------------
>
> Key: HBASE-2925
> URL: https://issues.apache.org/jira/browse/HBASE-2925
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.20.3, 0.90.0
> Reporter: Robert Mahfoud
> Attachments: 2925.txt, SimpleHConnectionManagerLeakReplicator.java
>
>
> {{HConnectionManager.getConnection(config)}} caches the created
> {{TableServer}} in {{HBASE_INSTANCES}} (a {{LinkedHashMap}} ) which is keyed
> by the configuration instance itself.
> Given the current implementation of {{hashCode()}} (and {{equals()}}) of
> {{HBaseConfiguration}}, the hash code of the configuration is changed if any
> of its properties are changed, which will cause the keys of
> {{HBASE_INSTANCES}} to be inconsistent with the hashtable that contains them,
> making some entries unreachable.
> In this case, when the map's LRU strategy needs to remove the oldest entry,
> it tries to remove it based on the oldest key, which no longer gives the
> original hash code, therefore the lookup in
> {{HBASE_INSTANCES.remove(oldest)}} doesn't actually remove anything.
> This has been observed to lead to OOM errors in long running clients.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
